Seebeck Effect in Magnetic Tunnel Junctions

Marvin Walter, Jakob Walowski, Vladyslav Zbarsky, Markus Münzenberg*

I. Physikalisches Institut, Georg-August-Universität Göttingen, Germany

Markus Schäfers, Daniel Ebke, Günter Reiss,

Department of Physics, Universität Bielefeld, Germany

Andy Thomas
Department of Physics, Universität Bielefeld, Germany and Institut für angewandte
Physik, Hamburg, Germany

Patrick Peretzki, Michael Seibt,

IV. Physikalisches Institut, Georg-August-Universität Göttingen, Germany

Jagadeesh S. Moodera, Massachusetts Institute of Technology (MIT),

Cambridge, USA

Michael Czerner, Michael Bachmann, Christian Heiliger, I. Physikalisches Institut,

Justus-Liebig-Universität Gießen, Germany

Abstract

Creating temperature gradients in magnetic nanostructures has resulted in a new research

direction, i.e., the combination of magneto- and thermoelectric effects [1,2,3,4,5]. Here, we
demonstrate the observation of one important effect of this class: the magneto-Seebeck effect.
It is observed when a magnetic configuration changes the charge based Seebeck coefficient. In
particular, the Seebeck coefficient changes during the transition from a parallel to an antiparallel
magnetic configuration in a tunnel junction. In that respect, it is the analog to the tunneling
magnetoresistance. The Seebeck coefficients in parallel and antiparallel configuration are in the
order of the voltages known from the charge-Seebeck effect. The size and sign of the effect can
be controlled by the composition of the electrodes' atomic layers adjacent to the barrier and the
temperature. The geometric center of the electronic density of states relative to the Fermi level
determines the size of the Seebeck effect. Experimentally, we realized 8.8 % magneto-Seebeck
effect, which results from a voltage change of about -8.7 µV/K from the antiparallel to the
parallel direction close to the predicted value of -12.1 µV/K. In contrast to the spin-Seebeck
effect it can be measured as a voltage change directly without conversion of a spin current.

*Corresponding author: Markus Münzenberg, e-mail: mmuenze@gwdg.de

The creation of an electric field by a temperature gradient in a material is known as Seebeck
effect since 1826. In the last years new spin dependent thermal effects have been discovered in
ferromagnets and the Seebeck effect receives novel interest. Gravier et. al. described the
transport of heat and spin in magnetic nanostructures [ 1]. Uchida et. al. experimentally found
the spin-Seebeck effect [ 2] driving this novel field, e.g., in nanoscale metal structures [ 3], in
magnetic insulators and semiconductors [ 4, 5]. A strong asymmetry of the density of states with
respect to the Fermi level promotes the heat-driven electron transport that leads to the common
charge-Seebeck effect. These strong asymmetries can be found in the spin split density of
states in ferromagnetic materials. Previously, the effect amplitude resulting from this spin
asymmetry was believed to be a second-order effect [ 6]. In this work, we demonstrate that the
magneto-Seebeck effect can be large. We first present ab initio calculations which show that
this effect can be of the same order as the charge-Seebeck effect. Using magnetic tunnel
junctions, where two ferromagnets are separated by a thin insulating tunnel barrier. The effect is
related to a half metallic behavior of the tunnel junction transmission. Our experiments show
that a thermoelectric power can be generated in such nanostructures over distances of only 2.1
nm, the thickness of the tunnel barrier. The change from parallel to antiparallel electrode
configuration is -8.7 µV/K at room temperature, while maintaining all other conditions in the
junction constant. Related with this magnetization switching we calculated a magneto-Seebeck
effect of 8.8 %. In theory, this change is predicted to be up to 1000 % / 100 µV/K [7]. In future
spincaloritronic [ 8] applications, the local cooling of an individual nanometer-sized area could,
therefore, be switched magnetically. The junction size allows stacking and nano-integration of
these novel thermopower devices.

The magneto-thermal effect is based on the seminal work conducted by Johnson and Silsbee
[ 9]. They gave a general description of the mechanisms that affect a ferromagnetic material
when a heat flow causes a temperature gradient. Strong thermomagnetic effects can be
expected in a half metal where the spin polarization can be up to 100% [10]. One can define a
spin-dependent Seebeck coefficient by replacing the charge-dependent Seebeck voltage by a
voltage generated for each spin channel. The difference between the two spin-dependent
Seebeck coefficients is driving a spin accumulation. In contrast, the magneto-Seebeck effect is
different from the spin-Seebeck effect, because it is not related to a spin-voltage generation. It
occurs in junctions and is similar to the giant (GMR) and tunneling magnetoresistance (TMR). It
results in a charge-Seebeck effect that is changed by the magnetic orientation of the electrodes.
This voltage is accessible directly without conversion. In order to have a high charge-Seebeck
effect, a high asymmetry in the energy dependence with respect to the electrochemical potential
for the transport states is necessary, realized in semiconductors as shown in Fig.1 a).
Consequently, for the thermomagnetic effect, these energy asymmetries must be different for
spin up and spin down carriers. For our experiments, the recent progress in giant tunneling
magnetoresistive (giant TMR) junctions enabled us to use magnetic tunnel junctions (MTJs) with
high spin asymmetry. Their large contrast in the spin-dependent transmission due to different
symmetries of the tunneling states in both spin channels should lead also to different energy
asymmetries of the tunneling states as shown in Fig. 1 b). We define the magneto-Seebeck
ratio (SMS) from the Seebeck coefficients in the parallel and anti-parallel configuration, as given
in equation (1):
 (1)

At first glance, it seems that the magnetoresistance, the spin-Seebeck effect, and the magneto-
Seebeck effect should be related to each other. However, these are different effects, and, in
general, it is not possible to calculate one from the others.

Figure 1 Origin of the magneto-Seebeck effect. a) Semiconductors are known to generate

high Seebeck effects. b) In magnetic tunnel junctions, thermal differences in the electron
distributions and strong asymmetry in the spin-dependent tunneling channels are depicted. T(E)
is the transmission of the full tunnel junction for which either the ferromagnetic electrodes can
be a highly spin polarized half metal or the combination of the barrier and the ferromagnet
exhibit half-metallic characteristics. The function is given in darker color.
The thick line marks the resulting value of the geometric center SP and SAP. In the lower
symmetric case, the magneto-Seebeck effect is vanishing. c) Calculation of the Seebeck
coefficients as a function of temperature for tunnel junctions with 10 monolayers of MgO as a
barrier. The magnetic layers are 20 monolayers thick. The semi-infinite leads are Cu in the bcc-
Fe structure. We assume a mixed termination of FeCo at the FeCo/MgO interface that is an
ordered, 2x1, in-plane supercell with one Fe and one Co atom. d) Seebeck coefficients for the
parallel (P) configuration and the antiparallel (AP) configuration are shown, e) is the
corresponding magneto-Seebeck effect SMS.

To understand this point, it is important to realize that the transport coefficients are calculated
from the transmission function T(E) of the tunnel junction but that they have different integral
values. The conductance g is determined by the integral of the transmission function T(E)
multiplied by the derivative of the electron occupation function ∂E f(E,μ,T) at temperature T and
electrochemical potential µ:

(2)

The Seebeck coefficient is also given by the transmission function T(E) multiplied by the
derivative of the occupation function ∂E f(E,μ,T):

(3)

Different to the magnetoresistance, the Seebeck coefficient is the geometric center of

. Fig. 1 b) illustrates these quantities for two different cases. The geometric
center for parallel and antiparallel configuration (SP and SAP) is marked by the thick line. We
assume a transmission function that has different energy asymmetries in both magnetic
configurations and different positions of the electrochemical potential. In the first case, a high
TMR and a high magneto-Seebeck ratio are obtained. In the second case, the value of SMS is
essentially zero, but the TMR is highest. Generally speaking, cases with vanishing value of SMS
and large TMR (or vice versa) are also possible. Therefore, we can tailor magnetic tunnel
junctions to be good candidates for large magneto-Seebeck effects. Consequently, we
investigated temperature-induced voltages in magnetic tunnel junctions starting with samples
showing large TMR ratios. Two different types of junctions with large TMR values could be
used, i.e., Fe-Co/MgO/Fe-Co and half-metallic compounds. We focus on the former case, since
it is demonstrated to have the largest experimental value of 604% at room temperature [ 11]. The
tunneling states of the electrons have been thoroughly investigated for MgO-based magnetic
tunnel junctions and the understanding of spin polarization of the current and the quantitative
approach to magnetoresistance in tunnel junctions has advanced enormously in recent years.
Our theoretical investigations are ab initio calculations based on density functional theory. In
particular, we used the Korringa-Kohn-Rostoker and the non-equilibrium Green’s function
method to obtain the transmission function T(E) [ 12]. Using T(E), we calculated the transport
coefficients according to Eqs. (2) and (3) [13, 14]. We investigated the magneto-Seebeck
coefficients for different temperatures for Fe0.5Co0.5/MgO/Fe0.5Co0.5 magnetic tunnel junctions
with bcc structure of the ferromagnetic electrodes. The temperature dependence is considered
only within the electron-occupation function. Due to coherent tunneling, the atomic structure of
the interface could be important. Therefore, we investigated the Seebeck coefficients for
different possible interface structures, i.e., the Fe-terminated structure, the Co-terminated
structure, and a mixed-termination structure. The results at a temperature of 300 K listed in
Table 1 show a strong dependence on the interface structure. Even a sign change was
observed. However, the case that the layer next to the barrier is pure Co or pure Fe is unlikely in
the experiment. Consequently, we continued our investigation with the mixed termination
structure (Co0.5Fe0.5). In Fig. 1 d), SP and SAP are plotted as a function of temperature for a tunnel
junction that has an MgO barrier that was 10 monolayers thick. In addition, we plot the
corresponding magneto-Seebeck ratios (Fig. 1 e). Although SP and SAP do not change sign, SMS
does when SP = SAP. We found that SP and SAP were large when compared to charge-Seebeck
coefficients.

Figure 2 Switching of the Seebeck effect via the magnetization. a) Schematics of the laser
heating setup and b) Au top contact geometry of the device with the laser spot dimensions. c)
From antiparallel d) to parallel orientation of the layer magnetization, the charge-Seebeck
voltage varies. By the magnitude of its change the magneto-Seebeck effect SMS is defined.

Table 1: The Seebeck coefficients for parallel SP and antiparallel configuration SAP and the
magneto-Seebeck effects calculated for different supercells at a temperature of 300 K. The
results show the sensitivity to the interface composition. SMS defines the relative change and
can be negative or positive.

FeCo/MgO/FeCo with a 10-monolayer MgO barrier

SP [μV/K] SAP [μV/K] SP - SAP [μV/K] SMS [%]
CoFe -19.7 -32.4 12.7 64.1
FeCo 45.9 -50.0 95.9 209.0
CFFC 9.4 -44.6 54.0 573.2
Co0.5Fe0.5 -34.0 -21.9 -12.1 -55.2
Experimental value -107.9 (-1300) -99.2 (-1195) -8.7 (-105) -8.8 (-8.8)

Abbreviations: CoFe - Co0.5Fe0.5 layers with Co at the MgO interface. FeCo - Co0.5Fe0.5 layers with Fe at the MgO
interface. CFFC - Co0.5Fe0.5 layers with Fe at one of the MgO interfaces and Co at the other. Co0.5Fe0.5 - super cell in-
plane with Co:Fe 1:1 at the interface. The values derived from the experiment are given for a temperature difference
at the MgO barrier of 53 mK (4.4 mK) respectively. The temperature difference ∆T is taken from the numerical
simulation of the temperature gradients using the thin film value (bulk value) of the thermal conductivity of MgO.

Figure 3 Cross sections and temperature gradients in the tunnel junction. a) TMR
junctions: device structure studied with Transmission Electron Microscopy (TEM). The high
resolution shows the epitaxial relationship Fe-Co(001)/MgO(001) for two transmission directions
MgO[100] and MgO[110]. b) On the right side, the simulated temperature distributions are
shown for 200 ps and 1 µs after the laser power of 30 mW is turned on in a two dimensional
cross section. c) The temperatures for the final static equilibrium condition are shown below as
a line scan.

For the experiments, we use Co-Fe-B/ MgO/ Co-Fe-B pseudo spin valve structures. The 1x1
μm2 tunnel junction is heated homogeneously by 30 mW laser power (diode laser with 15-20 μm
focus in diameter and a wavelength of 784 nm, Fig 2 a, b) and the charge-Seebeck voltage
(Seebeck voltage in the following) is measured for the parallel and antiparallel orientation of the
layer magnetization (Fig. 2 c, d). To obtain the temperature distribution and the time constants
for the heat diffusion, we used finite element simulations. Transmission Electron Microscopy
(TEM) in Fig. 3 a) reveals the device geometry that serves as an input to integrate the heat
diffusion equation. In order to calculate the Seebeck coefficients, we estimate a temperature
difference ∆T at the 2.1 nm MgO barrier. For polycrystalline MgO films with a nanometer grain
size, the heat conductance is lower than the bulk value due to the grain boundaries [15]. The
high resolution TEM in Fig. 3a, however, reveals a good crystalline quality of the investigated
samples. Nevertheless, the thermal resistance at the Co-Fe/MgO interfaces can have similar
effects as the grain boundaries. Therefore, we used both the bulk as well as the reduced value
for the thermal conductivities as given in the supplementary information.
In Fig. 3 b) we show the resulting temperature profile in a two dimensional cross section for 200
ps and 1 µs after the laser power is turned on. A series allows determining the time scales of the
heating: the static temperature profile is reached after about 2 µs. The final temperature
distribution is shown as a line scan in Fig. 3 c) across the tunnel junction. A temperature
difference at the 2.1 nm MgO barrier of 53 mK (4.4 mK) is derived from the numerical simulation
using the thin film value (bulk value) of the thermal conductivity of MgO respectively.

Figure 4 Seebeck voltages for Fe-Co-B/ MgO/ Fe-Co-B elements. a) Temporal traces are
shown for different modulation frequencies. The heating by the laser is shown in red overlaid to
the corresponding Seebeck voltage (the voltages are shifted for clarity). b) Magneto-Seebeck
voltages are given above for 1.5 and 3 kHz lock-in modulation. c) The corresponding
magnetoresistance (TMR) shows a hard-soft switching of the pseudo spin valve structure.

Fig. 4 shows the magneto-Seebeck effect of a single MTJ with a TMR of 150 %. The temporal
voltage traces in Fig. 4 a), as observed in several junctions, show a peak-like voltage when the
laser heating is increased and decreased periodically. A negative peak occurs when the laser
power is turned on. From the time constants simulated we identify this voltage peak with the
Seebeck voltage generated at the junction. The shutter moving through the laser spot limits the
time scale to approximately 10-100 µs in our data. As a reference the sequence of the
measured laser power is given for each signal trace. The voltage reverses sign when the laser
heating is turned off. The Au pads efficiently conduct the heat away from the heat spot
(extension of 17.5 µm) into the large bond pads that act as a heat sink. The other side of the
junction is then still at a higher temperature and the temperature gradient is reversed. In the
lowest curve with a 800 Hz modulation frequency, the magnetic tunnel junction was heated
asymmetrically allowing a longer cooling time. The Seebeck voltages for the parallel and
antiparallel configurations are determined from Fig. 4 b), in which the Seebeck voltage is shown
as a function of the applied field. As expected from the temporal traces, a larger value was
found for the fastest modulation of the laser power (at 3 kHz). The Seebeck voltage at the
junction contributes more to the total signal than it does for the slower modulation, where the
whole sample heats up on a larger area. A signal proportional to the modulation frequency is the
dominant component. One obtains -5.7 μV for the parallel and -5.3 µV for the antiparallel
orientation, i.e., a change of the Seebeck voltage by 0.4 µV for the tunnel junction. Fluence
dependent experiments suggest that the increase of the Seebeck voltage with laser power
depends on the increase of the temperature gradient at the barrier and the base temperature at
the junction that is increased by the laser power as well, which is discussed further in the
supplementary information. If the junction barrier is pushed through a dielectric breakdown [16],
the magneto-Seebeck effect disappears.

The experimental results and the theoretical predictions for the Seebeck coefficients are
summarized in Table 1. The theoretical prediction for the Co0.5Fe0.5 case with Co:Fe 1:1 at the
interface is closest to the experiment. The values are negative for both the parallel and the
antiparallel configuration. To calculate the Seebeck coefficients from the experimentally
determined Seebeck voltage VP,AP, we take the temperature gradient to be 53 mK across the 2.1
nm MgO tunnel barrier from our numerical modeling. Thus, one obtains a value of VP /∆T= -108
μV/K for the Seebeck coefficient SP for the parallel orientation. A decrease of the thermal
gradient taking the bulk value for the MgO thermal conductivity as input parameter in our model
increases the Seebeck coefficients calculated accordingly. This allows deriving an upper limit of
-1300 μV/K for the Seebeck coefficient, given in brackets. Note that spurious other voltages
generated in the layer stacks or within the heated device change the Seebeck effect amplitude,
but not the difference of the Seebeck voltage for parallel and antiparallel configuration. In
accordance with this, the experimental results for the difference (SP-SAP) of -8.7 μV/K are closer
to the predicted value of -12.1 μV/K than the individual values of SP and SAP. The lower limit of
the magneto-Seebeck effect, i.e., the relative change of the Seebeck voltage for the parallel and
the antiparallel case, is -8.8 %.

Finally, the magneto-Seebeck effect in magnetic tunnel junctions allows controlling these
effects. As a major strategy to develop the possibilities opening up with the magneto-Seebeck
effect it is crucial to tailor the thermal tunneling current arising from the majority and the minority
spins, i.e. to maximize the shift of the geometric center relative to the Fermi level of these
electronic states contributing to the thermal transport. The calculations demonstrate that even
the sign of the magneto-Seebeck effect can be controlled using different Co-Fe compositions.
Findings on tunnel junctions using a different method (resistive heating) yielding the same effect
magnitudes with different sign were recently reported [17] using Co-Fe-B/ MgO/ Co-Fe-B
devices however of different composition and structure (using a Singulus Tech. cluster tool).
The qualitative change for the devices presented in our letter however allows a comparison of
theoretical and experimental results. The results presented compare well to the theoretically
predicted change, including the predicted sign reversal of the magneto-Seebeck effect at
elevated temperatures (see supplementary information). Further, the experiments showed that
the magneto-Seebeck effect can be generated over length scales of only a few nanometers -
across a 2.1 nm thick tunnel barrier in our case. This reveals that the magneto-Seebeck effect in
magnetic tunnel junctions is a novel effect that can be used to manipulate and design
thermovoltages in nanometer-scale devices. The contrast for switching the voltage can be
increased further in the future which allows a control of the Seebeck effect by magnetic
switching.

Methods

Fabrication:

The Co-Fe-B films were prepared by magnetron sputtering using 2-inch targets with
compositions of Co0.4Fe0.4B0.2 (analysis Co:Fe 0.52:0.48) and Co0.2Fe0.6B0.2 (analysis Co:Fe
0.32:0.68) in a UHV system with a base pressure of 5x10-10 mbar. They are annealed ex-situ at
temperatures of 450°C up to 550°C (post growth annealing, 20-60 min). Samples prepared in
the Göttingen chamber were e-beam evaporated MgO after transferring in a separate UHV
chamber with base pressure of 5x10-10 mbar (maximum TMR reached is 200% at RT). With the
Bielefeld chamber, MgO was prepared by magnetron sputtering (maximum TMR reached is
330% at RT). The sample stack was reduced to a simple pseudo spin valve structure to
minimize the contribution of spurious Seebeck voltages at metal interfaces not stemming from
the junction: Au 27 nm/Ru 3 nm/Ta 5 nm/Co-Fe-B 5.4 nm/MgO 2.1 nm/Co-Fe-B 2.5 nm/Ta 5
nm/SiO2 500 nm/Si(100). This was done as a trade-off with the magnetic separation of the
switching fields, because no anti-ferromagnetic exchange bias layer is used which could also
give a magnetic contribution. After ex-situ annealing in a constant field, further structuring was
done by standard ion beam etching to yield 1x1 µm2 to 12.5x12.5 µm2 junctions to the MgO
barrier. The high resolution transmission electron microscopy (HR-TEM) data in Fig. 4 (bottom,
left) reveal the coherent growth of crystallized Co-Fe(110) at each side of the MgO(100) barrier
(solid state epitaxy) in columns that can be identified (MgO[001] and MgO[110] in the
transmission direction). As an isolation layer at the sides of the element, a 100-nm-thick SiO2
layer was grown by thermal evaporation. A 100 nm thick, top-contact Au layer was deposited as
a bond pad. This also prevents direct optical carrier excitation in the MgO barrier. A 5 nm layer
of Cr was deposited below the Au top-contact layer for better adhesion on the SiO2 isolation.

Experimental setup:

For the laser heating, a 100 mW, Toptica, intensity-stabilized laser diode module (λ = 784 nm)
was focused to a diameter of 15-20 μm full width at half maximum (FWHM). For the standard
experiments we used 30 mW laser power. The beam position on top of the bond pad was
controlled via a camera. The intensity was modulated using 800 Hz, 1.5 kHz, and 3 kHz
modulation frequency. To prevent a current flow in the system that could be modified by the
change of the resistance of the junction, a high input impedance (100 GΩ) LT1113 precision
operational amplifier (Linear Technology) was used. The band width of the amplifier is 5 MHz.
This was installed close to the sample to minimize the effect of the cable capacitance (< 15 pF).
A simulation of the circuit with the sample resistance showed that the change of the resistance
will contribute < 1 nV to the absolute voltage. Seebeck voltage versus magnetic field curves is
measured using a Stanford Linear research lock-in amplifier.

Thermal modeling of parameters (COMSOL):

To simulate heat flow and temperature distribution, the MTJ was modeled using the COMSOL
finite-element package. The tunnel junction geometry was taken from the cross sectional TEM
data as input parameters. The element was embedded into a 3 μm cylinder. The heat flow from
the laser heating comes from the top. For a 30 mW laser power the absorbed laser power is 10
mW. The temperature at the bottom of the cylinder in the Si(100) substrate was set to ambient
temperature. The 500 nm SiO2 layer on top of the substrate is the bottleneck for heat diffusion
through the cylinder stack. The temperature at the bottom layer of the element depends
sensitively on the heat flow though SiO2 layer and determines the 2 µs needed to reach the final
heat gradient. A prism-shaped, adaptive mesh was used with about >10 nm in the plane and
sub-nanometer resolution perpendicular to the plane. In addition, we performed simulations on
the larger length scale to simulate lateral heat diffusion. The heated area extends to about a
17.5 µm diameter, and, in that case, the absolute temperature increase is reduced to 8 K. The
equilibrium heat gradient is attained within about 2 µs. For the 2.1 nm MgO thin tunnel barrier, a
value of the thermal conductance of κ = 4 W/(m·K) is assumed to be closest to the reality. This
value has been determined experimentally for a thin film [15]. It is expected to be much closer to
the bulk value of κ = 48 W/(m·K) which gives an upper limit for the Seebeck coefficient. All
material parameters used in the numerical model are provided in a table in the supplementary
information.

References

[1] L. Gravier, S. Serrano-Guisan, F. Reuse, and J.P. Ansermet, Thermodynamic description

of heat and spin transport in magnetic nanostructures. Phys. Rev. B 73, 024419 (2006).
[2] K. Uchida, S. Takahashi, K. Harii, J. Ieda, W. Koshibae, K. Ando, S. Maekawa, E. Saitoh,
Observation of the spin- Seebeck effect. Nature 455, 778-781 (2008).
[3] A. Slachter, F.L. Bakker, J.P. Adam, and B.J. van Wees, Thermally driven spin injection
from a ferromagnet into a non-magnetic metal. Nature Phys. 6, 879–882 (2010).
[4] K. Uchida, J. Xiaso, H. Adachi, J. Ohe, S. Takahasi, J. Leda, T. Ota, Y. Kajiwara, H.
Umezawa, H. Kawai, G.E.W. Bauer, S. Maekawa, and E. Saitoh, Spin Seebeck insulator,
Nature Mater. 9, 894–897 (2010).
[5] C.M. Jaworski, J. Yang, D.D. Awschalom, J.P. Heremans, and R.C. Myers, Observation of
the spin-dependent Seebeck effect in a ferromagnetic semiconductor. Nature Mater. 9,
898–903 (2010).
[6] Typically the spin-orbit interaction relevant for the magnetic elements in the spin diffusion
tensor is a small parameter. However, here, the exchange interaction is the energy scale
imposing the spin asymmetries in the barrier transmission.
[7] M. Czerner, M. Bachmann, and C. Heiliger, Spin caloritronics in magnetic tunnel junctions:
Ab initio studies, Phys. Rev. B, Phys. Rev. B 83, 132405 (2011).
[8] G.E.W. Bauer, A.H. MacDonald, and S. Maekawa, Spin Caloritronics, Solid State Comm.
150, 459-460 (2010).
[9] M. Johnson and R. H. Silsbee, Thermodynamic analysis of interfacial transport and of the
thermomagnetoelectric system, Phys. Rev. B 35, 4959–4972 (1987).
[10] G. Müller, J. Walowski, M. Djordjevic, G.-X. Miao, A. Gupta, A. V. Ramos, K. Gehrke, V.
Moshnyaga, K. Samwer, J. Schmalhorst, A. Thomas, A. Hütten, G. Reiss, J. S. Moodera,
M. Münzenberg, Spin polarization in half metals probed by femtosecond spin excitation
Nature Mater. 8, 56-61 (2009).
[11] S. Ikeda, J. Hayakawa, Y. Ashizawa, Y. M. Lee, K. Miura, H. Hasegawa, M. Tsunoda, F.
Matsukura, and H. Ohno, Tunnel magnetoresistance of 604% at 300 K by suppression of
Ta diffusion in CoFeB/MgO/CoFeB pseudo-spin-valves annealed at high temperature,
Appl. Phys. Lett. 93, 082508 (2008).
[12] C. Heiliger, M. Czerner, B. Yu, Yavorsky, I. Mertig, M. D. Stiles, Implementation of a Non-
Equilibrium Green’s Function Method to calculate Spin Transfer Torque., J. Appl. Phys.
103, 07A709 (2008).
[13] U. Sivan and Y. Imry, Multichannel Landauer formula for thermoelectric transport with
application to thermopower near the mobility edge, Phys. Rev. B 33, 551-558 (1986).
[14] Y. Ouyang and J. Guoa, A theoretical study on thermoelectric properties of graphene
nanoribbons, Appl. Phys. Lett. 94, 263107 (2009).
[15] S.M. Lee, D.G. Cahill, and T.H. Allen, Thermal conductivity of sputtered oxide films. Phys.
Rev. B 52, 253-257 (1995).
[16 ] A. Thomas, V. Drewello, M. Schäfers, A. Weddemann, G. Reiss, G. Eilers, M.
Münzenberg, K. Thiel, M. Seibt, Direct imaging of the structural change generated by
dielectric breakdown in MgO based magnetic tunnel junctions, Appl. Phys. Lett. 93 ,
152508 (2008).
[17] N. Liebing, S. Serrano-Guisan, K. Rott, G. Reiss, J. Langer, B. Ocker and H. W.
Schumacher, Tunnelling magneto thermo power in magnetic tunnel junction nanopillars,
arXiv 1104.0537.

Author contributions

M.W. and J.W. performed experiments; M.W., V.Z., M.Sch., and D.E. characterized and
prepared the TMR devices; P.P. and M.S. performed the high resolution TEM; M.W., J.W., and
M.M. analyzed the data; M.W. performed the COMSOL calculations; M.C., M.B., and C.H. did
the ab-initio transport calculations; A.T. and M.M. designed the research approach; C.H., M.M.,
and A.T. wrote the manuscript and developed the model; M.W., J.S.M, A.T., M.M., and C.H.
contributed to the development of the experiments; G.R., J.S.M, A.T., M.M., and C.H. and all
authors discussed the experiments and the manuscript.

Acknowledgments

A.T. acknowledges the Ministry of Innovation, Science and Research (MIWF) of the North
Rhine-Westphalia state government for financial support. M.M., M.S., M.W., and P.P.
acknowledge the funding provided by the German research foundation (DFG) via the SFB 602
for the TEM work. M.C., M.B., and C.H. acknowledge support from DFG SPP 1386 and DFG
grant HE 5922/1-1. J.S.M acknowledges support by the US National Science Foundation (NSF)
and Office of Naval Research (ONR). We acknowledge Anissa Zeghuzi’s help for additional
COMSOL calculations. This work was initiated by the SpinCaT priority program.
Seebeck Effect in Magnetic Tunnel Junctions

Supplementary information

1. Details of the laser heating and Seebeck voltage transients

The slowest time constant in the heating process arises from the low conducting electric
isolation SiO2 layer on the Si wafer, which is the thermal bottleneck. The temperature
gradient at the barrier itself, which is responsible for the voltage generation, is 4.4 mK, in
accordance with the simple estimation using the ideal bulk value for MgO for the thermal
conductivity κ. A closer fit of theoretical and experimental results will be possible with a more
precise experimental determination of the temperature gradient at the MgO barrier, since it
depends sensitively on the thermal conductivity of the thin films and interfaces and thus has
the largest error bar. While, in principle, this task is possible for such thin films [1], it will be
difficult to accomplish especially in a single device. However the thin film value for the 2.1 nm
MgO layer is expected to be less than its bulk value of κ = 48 W/(m·K). For polycrystalline
MgO films with a grain size of 3-7 nm, a value of κ = 4 W/(m·K) was found at room
temperature, which differs by a factor of ten [2]. This difference arises from the grain
boundaries. In addition, the thermal resistance at the Co-Fe/MgO interface can be
considerable, i.e., 1x109 W/(m2·K) [1]. This interface term is approximately of the same order
as the thermal resistance of a film that is 1 nm thick and may further increase the thermal
gradient and thus decrease the Seebeck coefficients calculated. Taking into account the
crystalline quality of the tunnel junctions grown by solid-state epitaxy as derived from the
high resolution TEM, the most realistic scenario is a model assuming bulk-like thermal
conductivity for the MgO because of its crystalline quality and a thermal interface resistance
at both Co-Fe/MgO interfaces. Thus the value for the MgO tunnel barrier is expected to be
between the thin film value for the polycrystalline material and the bulk value. Here we give
the polycrystalline and the bulk value as a lower (conservative estimation) and upper limit
(optimistic estimation).

Material κ [W/(m·K)] ρ [103 kg/m3] cv/cp [J/(kg·K)]

Ta 57 16.7 140
Ru 117 12.4 238
Au 320 19.3 128
Cr 94 7.2 449
MgO 4 (48) 3.6 935
SiO2 1.4 2.2 1052
Co20Fe60B20 87 8.2 440

Table 1: For MgO the thin film value (bulk value) of the thermal conductivity given respectively.

In the following, we discuss the dynamic time scales that contribute to the voltage response
to the laser heating experiments shown for different modulation frequencies in Fig. 4a (main
manuscript). Time traces have been measured after amplification (100-MHz PM 3382 Phillips
oscilloscope). A negative voltage was found for the opening of the shutter upon heating.
When the shutter is closed, the cooling reverses the sign of the voltage. The heat is
effectively conducted away from the 17.5 µm heat spot into the large 100 nm thick Au bond
pads. The time constant of the voltage increases within a few µs, i.e., the response of the
voltages are approximately 10-100 µs, generally asymmetric for the heating and cooling
process. The dynamic response depends on three factors, all of which were in the
microsecond regime, i.e., 1) the time constant of heating, 2) the width of the Gaussian laser
intensity profile that moves across the chopper blade, and 3) the response time of the
LT1113 precision operational amplifier to realize a high input impedance of 100 GΩ. While
the thermal equilibration time in the COMSOL model is 2 µs, and the response time of the
amplifier 5 MHz, we identified the width of the Gaussian laser intensity profile that moves
across the chopper blade as slowest contribution limiting the response time. The laser
diameter and chopper blade dimensions (different for each modulation frequency with varied
chopper blades) range from 20-150 µs and can be extracted from the reference diode signal
in Fig. 4a (main manuscript).

In addition, we conducted tests to determine whether spurious magneto voltages were

generated in the layer stack. A natural failure of the barrier (dielectric breakdown) at low
voltages allows shorten the junction without too much destruction of the element and the
magneto-Seebeck effect should vanish. To prove this claim, we conducted tests to determine
whether magneto-Seebeck voltage was generated at the tunnel barrier. The tests consisted
of gradually shorting the barrier, as evidenced by a decrease in the resistance from the 100
kΩ of an intact junction to 1 kΩ and, finally, to 100 Ω for the shortened junction. In the
intermediate state, the magneto-Seebeck effect decreased to about 1%, but the magnetic
effect in the Seebeck voltage vanishes completely if the MgO tunnel barrier is bridged. This
excludes the existence of magneto voltages not generated at the tunnel barrier. However, a
background voltage is still contributing to the signal, thereby limiting the experimentally
determined value of the magneto-Seebeck effect, SMS. Also, we can exclude the generation
of lateral spin currents. This was checked by reversing the direction of rotation of the chopper
blade and thereby changing the direction of the possible lateral heat gradients in plane. This
sign change should in that case also lead to sign reversal which is not observed.

[1] David G. Cahill, Analysis of heat flow in layered structures for time-domain
thermoreflectance. Rev. Sci. Instrum. 75, 5119 (2004).
[2] S.M. Lee and D.G. Cahill, T.H. Allen, Thermal conductivity of sputtered oxide films.
Phys. Rev. B 52, 253-257 (1995).

Fig. 1: COMSOL simulation on the larger length scale to illustrate the extension of the
heating spot of 15 µm. The diameter of the spot heated up extends to 17.5 µm.
 2. Lateral heating in the Au bond pad

Shown in Fig. 1 is the full bond pad that has been simulated to reveal the extension of the
heated area. The simulation shows that the Au bond pad is not heated homogeneously. The
Gaussian beam diameter of 15 µm results in a heated area of 17.5 µm in diameter, not much
further extended than the area heated by the laser spot. The reason is the large heat
conductivity of 320 W/(m·K) of the 100 nm thick Au bond pad layer. During one modulation
cycle first, the laser power is turned on the temperature rises. Second, when turned off, the
large Au bond pads that have the highest heat conductivity work as a cooler. The
temperature gradient is reversed in the junction. This is detected in the sign reversal of the
Seebeck voltage for the transient voltage traces.

Fig. 2: Seebeck voltage measured for a tunnel junction heated with low laser fluencies (diode
laser). a) Seebeck voltage versus applied magnetic field. b) Seebeck voltages for parallel
and antiparallel orientation. The model is derived assuming a change of the Seebeck
coefficient with temperature.

3. Fluence dependence of the Seebeck voltages

In the following the fluence dependence of the Seebeck voltage is discussed. The
magnetoresistance of the junction is 120 % and it shows a magneto-Seebeck effect of 6.2 %,
which is somewhat smaller than for the element discussed in the main manuscript in
particular. The voltage change from parallel to antiparallel configuration is 2.5 µV (for 30 mW
laser power). In the experiment the laser fluence is varied from 5 to 40 mW. It can be seen
that with increasing the laser fluence the shape of the signal does not change. However the
Seebeck voltage increases as expected. For the highest laser fluence one finds -45 µV in the
parallel configuration and -42 µV for the antiparallel configuration. To discuss the evolution in
detail, the values are given in Fig. 2. One observes that the voltage increases nonlinearly
with the pump fluence. However this behavior is expected: with increasing the pump-fluence
not only the temperature difference at the barrier increases, also the base temperature
increases. One can calculate the voltage dependence from the following model

VP , AP = S P , AP ∆T
(1)
= s P , APT a ( T − 293K )

Here sP,AP is defined as the slope of the SP,AP versus T curve. The parameter a is given by
the total junction geometry. It relates the temperature gradient at the junction (∆T) to the
temperature gradient in between the tunnel barrier (T) and room temperature (293 K) at the
substrate bottom. From a series of simulations using laser fluences from 5 to 40 mW, we
estimate a value of 1.8 mK/ mW as rise in the temperature gradient ∆T with laser power. The
increase of the temperature at the tunnel barrier is 1.2 K/ mW with laser power. In addition,
the theoretical model predicts that the Seebeck coefficient itself decreases as the base
temperature is raised. Assuming a constant slope sP,AP, we can qualitatively understand the
linear plus quadratic behavior of the data in Fig. 2b.

Fig. 3: Seebeck voltage measured for a tunnel junction heated with high laser fluences
(Ti:sapphire oscillator, central wave length λ=810 nm). a) Seebeck voltage versus applied
magnetic field. b) Magneto-Seebeck effect as a function of the laser power and
corresponding temperature calculated by the numerical model. The data for the diode laser is
shown for direct comparison. A sign change is observed in the region at around 400 K as
expected from the theoretical model.

Even larger fluences are reached using a Ti:sapphire laser, which can be regarded here as a
cw laser source at moment. This is justified to a good approximation because of the large
distance from the Au top layer to the tunnel barrier (>130nm). We find that the magneto-
Seebeck values for high laser fluences lie on top of the previous data determined using the
cw diode laser, both shown in the shown in the same figure for direct comparison (Fig. 3b).
One observes a characteristic sign change of the magneto-Seebeck effect in-between 60
and 90 mW. This sign change is predicted our theoretical model (Fig. 1 e, main manuscript).
By relating the pump fluence to a temperature using the numerical finite element simulation,
one can plot the magneto-Seebeck effect as a function of the temperature derived at the
tunnel junction (Fig. 3b). For the largest laser power the base temperature can be increased
up to 475 K. The sign reversal takes place at around the predicted compensation point for
the Seebeck voltages for parallel and antiparallel orientation at 400 K. However there are
also features observed not expected from the model. The magneto-Seebeck effect shows a
decrease in between 300 to 350 K before it rises and crosses zero. However, the magneto-
Seebeck effect reaches 40% for 150 mW laser power, close to the theoretically expected
value. The maximum voltage change from parallel to antiparallel configuration is 15 µV.
Altogether the finding of the prominent feature, the sign reversal with increased temperature,
supports the conclusion that the transport coefficients calculated according to the model of
coherent tunneling model can explain characteristic features and thus is expected to be the
dominating contribution to the magneto-Seebeck effect.

In the high fluence range one may ask if for such large temperature gradients the theoretical
model is still valid. The temperature gradient at the 2.1 nm thick MgO barrier is 270 mK for
the highest fluence. To answer this question, it is helpful to compare this temperature to an
energy scale. Bias dependence is calculated in such system for up to 1 V bias voltage [3].
This amounts to 1 eV disturbance of the electrochemical potential or huge voltage gradients
of 10.000.000 V/cm. Therefore the temperature gradients here can be regarded as a much
weaker disturbance.

[3] C. Heiliger, P. Zahn, B. Y. Yavorsky, I. Mertig, Interface structure and bias

dependence of Fe/MgO/Fe tunnel junctions: Ab initio calculations, Phys. Rev. B 73,
214441 (2006).